iiinova — where your digital self keeps evolving

Grumpy Monoceros @grumpymonoceros.bsky.social · 26/09/2026

Being able to empathize with people who believe different things than you, and find common ground where it does exist, are fundamental political skills that we've just...decided not to exercise for some reason.

000

Grumpy Monoceros @grumpymonoceros.bsky.social · 26/09/2026

And there's a great deal of infighting and disagreement amongst different factions of liberals who are 90% in alignment on most things because the idea of banding together with someone who isn't in complete alignment is anathema in our political culture.

101

Grumpy Monoceros @grumpymonoceros.bsky.social · 26/09/2026

Even Republicans, as much as they've all collectively agreed to sell out their values and embrace their worst impulses, contain different groups with different goals and beliefs (understanding those beliefs is just more complicated than listening to what they say)

100

Grumpy Monoceros @grumpymonoceros.bsky.social · 26/09/2026

Who is "the other side" here? At any given moment there's a great deal of disagreement between lots of different groups, and not all of them are Republicans.

111

Grumpy Monoceros @grumpymonoceros.bsky.social · 26/09/2026

You know when there's a misunderstanding owing to the fact that we're all talking with strangers about complex things in 300 character chunks one could just...clear up the miscommunication. Nothing she said was particularly hostile or accusatory

110

Grumpy Monoceros @grumpymonoceros.bsky.social · 26/09/2026

Sure I was responding to the more recent conversation. A star trek future where professional computer touchers are unnecessary sounds rad

000

Grumpy Monoceros @grumpymonoceros.bsky.social · 26/09/2026

To put this another way: I think very highly of my ability to solve problems, but no one should ask me to diagnose them with a disease because I lack the fundamental medical training to solve *medical* problems.

010

Grumpy Monoceros @grumpymonoceros.bsky.social · 26/09/2026

can AI maybe help bridge that gap, at least for certain personal software type use cases? Sure, *especially* if that person is tech savvy. But the useful skillset of a programmer isn't actually the knowing of programming language syntax.

300

Grumpy Monoceros @grumpymonoceros.bsky.social · 26/09/2026

It's not really the capacity for problem solving, it's education in how computers work. When you get a CS degree you're not learning a bunch of programming languages you're learning algorithms and protocols and the fundamental architecture of how software functions.

320

Grumpy Monoceros @grumpymonoceros.bsky.social · 24/09/2026

And remember, they made these choices way before the era where autonomous agents were feasible. These choices were made in an era where they really did just spit out text - Google famously had a superbowl ad telling you Gemini was for helping your daughter write fan letters to famous athletes!

010

Grumpy Monoceros @grumpymonoceros.bsky.social · 24/09/2026

There are a million terms you could have used to describe these tools (like "LLMs" or the suggested "probabilistic automation") and they chose the one that obscures the mechanisms of how they work.

111

Grumpy Monoceros @grumpymonoceros.bsky.social · 24/09/2026

And I do think this was an intentional choice to mislead by the AI labs. The term "AI" is technically correct in a CS jargon sense but it's extremely nonspecific.

110

Grumpy Monoceros @grumpymonoceros.bsky.social · 24/09/2026

I think you're significantly underestimating the degree to which people understand computers at all. There are people out there who were driven to psychosis and suicide because they talked to *and trusted* an LLM - those people don't have a nuanced understanding of how LLM "thinking" is different

120

Grumpy Monoceros @grumpymonoceros.bsky.social · 24/09/2026

Through a million tiny little decisions the AI companies actively encouraged us to anthropomorphize them, down to the decision to deploy them publicly as chatbots.

000

Grumpy Monoceros @grumpymonoceros.bsky.social · 24/09/2026

I also think to some extent *any* attempt to change how we talk about LLMs will intrinsically feel obtuse, just because we've been using anthropomorphizing language to describe them for years now.

100

Grumpy Monoceros @grumpymonoceros.bsky.social · 24/09/2026

2 reads really obtusely because it is not clear what "trained software" is or why it would go off and break stuff like that. 1 is a little nonspecific but I don't see any essential facts missing. My main critique is that it could describe LLM generation as probabilistic/nondeterministic.

230

Grumpy Monoceros @grumpymonoceros.bsky.social · 24/09/2026

By contrast, anthropomorphized explanations require the viewer to already have a working understanding of how deep learning works which is not something we should expect of the average person.

100

Grumpy Monoceros @grumpymonoceros.bsky.social · 24/09/2026

2-4 are jagon-y and difficult to parse without knowing the lingo, but #1 seems pretty bang on to me. It's an oversimplification of what an LLM does, sure, but any explanation that a layman could understand would be, and that is close enough.

2100

Grumpy Monoceros @grumpymonoceros.bsky.social · 24/09/2026

But the core failure mode being described here - that an LLM running autonomously will lose the plot and start building the wrong thing - has been pretty consistent throughout.

010

Grumpy Monoceros @grumpymonoceros.bsky.social · 24/09/2026

Honestly I don't think my overall experience with Opus has changed that much since the release of Opus 4. Over time it gets a little more clever, it gets a little better at finding the right information on its own, it's better at testing/validation unprompted.

100

Grumpy Monoceros @grumpymonoceros.bsky.social · 23/09/2026

But systems are (at least in a functional political environment) very changeable. If the problem is just the systems that surround AI we can regulate the change that needs to happen (and maybe exile Sam Altman to the moon for funsies)

110

Grumpy Monoceros @grumpymonoceros.bsky.social · 23/09/2026

I think it's worth separating the systems around the technology from the technology itself. Framing this as a problem with the technology limits your options on how to deal with it. Technology is really just knowledge and infrastructure, if technology has a flaw finding a solution is hard.

110

Grumpy Monoceros @grumpymonoceros.bsky.social · 23/09/2026

It's going in a thousand different directions at once and some of them are pretty Star-Trek-y. Being able to just tell your computer to do something and then have it do that thing is very Star Trek. Tillie Norwood is...not

110

Grumpy Monoceros @grumpymonoceros.bsky.social · 23/09/2026

But the arts don't really work the same way. They're less about the economic value and more about personal expression (& the value to society *is* in large part perceived to be the personal expression)

100

Grumpy Monoceros @grumpymonoceros.bsky.social · 23/09/2026

I think this is correct. I view automation of economically valuable work as a moral imperative and as much as I don't want to have to figure out a new job it'd probably be good for society for technology to work like Star Trek (no need for professional computer touchers acting as gatekeepers)

100

Grumpy Monoceros @grumpymonoceros.bsky.social · 22/09/2026

If you were right and this is just a case of the author being a shit parent, then what does that imply about the school? Are all the poor Black schools just filled to the brim with shit parents for some reason? If not, how can you blame the author's parenting and not blame the parenting of the rest?

010

Grumpy Monoceros @grumpymonoceros.bsky.social · 22/09/2026

It's also worth checking on how your interpretation jives with the broader statistics. We know that the author's daughter was a high-performing student at her school, but her school overall performed terribly.

110

Grumpy Monoceros @grumpymonoceros.bsky.social · 22/09/2026

The author, a professional writer, had a point to this article, and that point wasn't "prove to the world I'm a good mom". The article would not be improved by an exhaustive retelling of every academic milestone as she saw it.

110

Grumpy Monoceros @grumpymonoceros.bsky.social · 22/09/2026

...But we also have no reason to know she didn't do that! The article doesn't mention her identifying her daughter's reading issues earlier on, but if she had, it would have added bloat to an already long article.

110

Grumpy Monoceros @grumpymonoceros.bsky.social · 22/09/2026

Your standard appears to be that this parent should have been able to read the non-school writing of her 4th grader and immediately clock whether or not her errors were grade-level-appropriate, which is an insane standard to hold parents to.

110

Grumpy Monoceros @grumpymonoceros.bsky.social · 22/09/2026

A 4th grader having writing assignments with errors is not the same as "can't write". Most 4th graders can write, but do not in fact write perfectly or with consistently correct spelling and grammar.

110

Grumpy Monoceros @grumpymonoceros.bsky.social · 22/09/2026

There's a section where it mentions she was below grade level on reading...but that was just because she had never taken the literacy assessment and in fact read just fine.

Snippet from the New York Times article

Then one day in December, Najya came home and excitedly informed me that she had suddenly jumped three grade levels in reading in one month. Knowing that this was virtually impossible, I reached out to the school’s literacy coach and learned that after we began asking for the reading assessment, Najya’s teacher had instructed the coach to perform the evaluation and that she found Najya was reading just fine. What became clear to me was that Najya’s teacher had most likely never performed the school’s mandatory reading assessments for Najya or any of her classmates but instead had assumed the flagging competency of these mostly low-income Black children.

030

Grumpy Monoceros @grumpymonoceros.bsky.social · 22/09/2026

Where are you getting that she can't write? Both examples in the article of her bad academic performance relate to math, and her earlier schooling famously didn't have homework so it's pretty easy for me to see how that slipped under the radar.

230

Grumpy Monoceros @grumpymonoceros.bsky.social · 22/09/2026

I am really curious what part of the article gave you this impression because that's not anywhere I can find, and in fact the daughter's school performance over time is discussed in significant detail.

260

Grumpy Monoceros @grumpymonoceros.bsky.social · 18/09/2026

Or, alternately, instead of just composing multiple models together: composing models with traditional deterministic software tricks.

010

Grumpy Monoceros @grumpymonoceros.bsky.social · 18/09/2026

I think a lot of the allure of LLMs is that they offer one model to 'do it all' but that comes with significant downsides (cost, latency, accuracy to name a couple). Jev IMO makes a strong argument that for many things we're better off with composing other more specialized models together.

110

Grumpy Monoceros @grumpymonoceros.bsky.social · 18/09/2026

Computer vision is also a thing that we've long been able to do quite well. Being able to look at an image and say with some confidence what it is isn't new to the LLM era.

110

Grumpy Monoceros @grumpymonoceros.bsky.social · 17/09/2026

LLMs being built at a time of peak scam culture sure is something

020

Grumpy Monoceros @grumpymonoceros.bsky.social · 17/09/2026

There's a tendency in tech (and among nontech people talking about tech) to assume that the tech is the only important skillset when often our work touches a huge number of other disciplines we're mostly not experts in. Academia figured this out but SV never did.

0122

Grumpy Monoceros @grumpymonoceros.bsky.social · 12/09/2026

Not sure if other harnesses have a similar option but I assume they all will at some point. This is pretty important

001

Grumpy Monoceros @grumpymonoceros.bsky.social · 10/09/2026

Welp learned about the TESCREAL bundle for the first time today and now I realize my prior contempt for SV billionaires was somehow insufficient. Thanks @timnitgebru.blacksky.app and @xriskology.bsky.social

firstmonday.org

View of The TESCREAL bundle: Eugenics and the promise of utopia through artificial general intelligence | First Monday

110

Grumpy Monoceros @grumpymonoceros.bsky.social · 09/09/2026

doesn't Anthropic publicly argue for regulatory scrutiny?

000

Grumpy Monoceros @grumpymonoceros.bsky.social · 09/09/2026

I'm probably not explaining this as well as I could (there's serious researchers who study alignment of which I am not one) but basically LLMs are a lot *more* predictable for short one-off tasks than they are for long-horizon tasks.

010

Grumpy Monoceros @grumpymonoceros.bsky.social · 09/09/2026

You're presupposing a loop where an agent is directed to do one very specific thing and then immediately report back to a person for further instructions. That is not how these agent systems work at all - they just keep working on it forever.

100

Grumpy Monoceros @grumpymonoceros.bsky.social · 09/09/2026

OpenAI tasked their model to solve a cybersecurity puzzle. Why did it go off and hack Huggingface's servers? Asking Claude implies you a human are in the loop and providing a short, directed task. In that case, it probably wouldn't, but that's not what anyone is talking about here.

100

Grumpy Monoceros @grumpymonoceros.bsky.social · 09/09/2026

Is there a particular reason you think they are not plausible?

100

Grumpy Monoceros @grumpymonoceros.bsky.social · 09/09/2026

Could be because a bad actor wanted it to. Could be because the agent was attempting to accomplish something else and figured this was the best way to do it. Could even have been accidental (it tried to do something else but ended up stealing the drone armada instead). All of these are plausible.

100

Grumpy Monoceros @grumpymonoceros.bsky.social · 09/09/2026

Because the drone armada is run by remote software that may have vulnerabilities that allow an agent to gain access it's not supposed to have. Drones are computerized systems that interact with the physical world, and so any system that gains control of them can interact with the physical world.

100